OK. Tried the music separator tool call "spleeter" to benchmark GPU tensorflow speed over CPU.
First tried it on "বাবু খাইছো?" downloading it form youtube.
CPU : 20 secs.
GPU : 15 secs.
Disappointed. Was hoping something around 10x speed improvement. That wasn't there.
Next "বিনোদিনী রায়"
CPU : 20 secs.
GPU : 18 secs.
Heart breaking. Maybe the extra money and time spent to get a GPU wasn't worth it.
Last wanted to check it on some long music, like 10 mins. Maybe that would make a different.
Tried "আগে কি সুন্দর"
CPU : 20 secs and crashed due to memory out.
GPU : 35 secs and finished successfully.
Even though GPU has only 2 GB ram and the system 8 GB. That was promising.
Lesson :
GPUs aren't significantly faster than pure CPU tensor. Rather invest on rake server if you indeed want to spend money on it.
Read a review somewhere that you don't need a GPU for ML as CPU speed is almost identical. But needed to see first hand to believe.
BUT,
You get stable result with GPU. With less RAM.
Shall test more and inform. InshaAllah.